Written by: Henry Lindsay-Smith, Head of Data Solutions,
and Joshua Dunn, Data Engineer, Mantel
What is the AI-enabled Data Delivery Lifecycle, and why is it important?
The data delivery lifecycle (DDLC) is a newer term in the industry than the well-established software delivery lifecycle (SDLC). Like the SDLC, the DDLC is changing rapidly as frontier model capabilities advance. To determine where AI can accelerate the DDLC, we first need a common definition of the DDLC.
We define the DDLC to be the process by which data teams deliver insights to their business stakeholders. This process moves from insights design to operation, as shown in the diagram below.
Each step traditionally ends in a labour-intensive formal document (e.g., a data-mapping Excel doc). The process itself is also linear. The problems with the DDLC are the labour-intensive manual processes and the overall linearity. This often manifests as complaints that it takes months to get a change to a report available or a new data product built.
How AI Accelerates the Data Delivery Lifecycle
Fundamentally, as (AI-enabled) data practitioners, we still try to achieve the same outcome: take business needs and deliver accurate data to enable stakeholders to make decisions. Each DDLC step still exists to some extent within the AI DDLC, but how we achieve it changes and, crucially, accelerates significantly. The key to this acceleration is giving the powerful frontier models the correct context and a clear goal.
We see the following innovations coming together.
- Workshops can be automatically transcribed to create requirements, and these requirements are captured in Markdown for the framework.
- The requirements Markdown + Data model + data modelling standards are consumed to build a new data model extension, with corresponding mapping files generated.
- DBT job built to populate the design model, with tests automatically created for guardrails
- A rapid iteration around model -> build -> test
The traditional linear flow above is replaced with an iterative loop that can bring timelines down from weeks to days.
Why Snowflake CoCo Excels in Data Delivery
So far, so good, but all of the above applies to any leading agentic coding harness. What makes CoCo special?
Snowflake has benchmarks on CoCo’s efficiency and a setup of bundled skills targeted to data engineers (see below), but for us, it is the simplicity. CoCo has access to the data in Snowflake and Snowflake’s current state by default. It is ‘Data Aware ’, not just ‘Code Aware’. You can achieve this on other platforms with other harnesses via MCP, but in our experience, having to set it up creates a productivity barrier.
With CoCo, from the moment you enable it, you can build and operate in the most efficient and accurate way. In typical Snowflake fashion ‘it just works’.
”Documentation is now almost unnecessary because CoCo can generate everything you need to know about your setup on the fly whenever you need to see the current state. It can help you debug everything, e.g. EntraID connectivity (check logs for successful logon). It can also propose a design for a data contract. Truly amazing.
Tim JelliffeLead Data Engineer, Mantel
Practical Snowflake CoCo features for data teams
As a tangible example from a recent project Mantel completed, the AI DDLC enabled by CoCo empowered us to build a data platform from the ground up, taking it from an empty account to a working platform with live use cases in less than 4 months. The accelerated development was largely due to CoCo and its ability to query data, write SQL, explore the Snowflake account, and more.
CoCo leveraged Snowflake’s native object discovery and natural language schema search, so the agent could inspect tables, read column metadata, and check existing models through natural-language prompts without the user needing to paste additional context such as query results or exact SQL table names.
CoCo also includes pre-packaged skill bundles that provide documentation for up-to-date Snowflake features in a digestible way for the agent. This resulted in consistently accurate outputs.
One of its most valuable capabilities is its customisability. For example, it can develop custom skills that provide tailored, repeatable workflows for an organisation or team.
Using Project-Scoped Skills in Snowflake CoCo
CoCo’s bundled skills cover Snowflake well, but they don’t know an organisation’s specific layer architecture, dbt macro conventions, or the specific intricacies of various source systems. Below are some examples of custom skills which immensely helped speed up development and the quality of the outputs on our project.
- dbt-transformation-pattern — Provided information on specific layer flow, what belongs at each layer, column naming standards, and model contracts. This also stopped the agent from putting business logic in bronze or entire layers in the architecture straight into gold.
- generate-curated-models — A skill backed by a Python script that scaffolds the full model chain from source metadata to develop consistent and standardised bronze and silver layer models, plus matching .yml files with tests. What previously took half a day of manual file creation and cross-referencing now takes minutes with a single prompt.
- dbt-sql-formatting — The Formatter. SQLFluff rules plus our project conventions (trailing commas, WHERE TRUE idiom, Jinja whitespace control), all wrapped up in one reference applied automatically during generation and explicitly during review.
The payoffs included both consistency and speed. Two engineers developing the same entity got structurally identical models, CTE shapes, test coverage, and header formats.
The Future of the AI Data Delivery Lifecycle with CoCo
The skills and features available through CoCo continue to evolve and offer capabilities beyond just using skills.
Skills encode what the agent should do. But on longer tasks involving multi-model generation, cross-layer refactors, and full sprint deliverables, the agent can still drift. This can lead to context loss due to compaction, approving its own work, or quietly inventing column names it hasn’t verified.
The next evolution pairs skills with subagents (purpose-built for a specific role) and hooks (event-driven scripts that fire regardless of model compliance). The open-source XO plugin is the clearest example. It combines one routing skill, four subagents with asymmetric tool grants (a writer that can’t explore, a critic that can’t edit), and three hooks that preserve state across compaction. The author never reviews its own output. The critic can’t modify what it reviews. Separate agents enforce structural constraints.
The tradeoff between the two approaches is described below.
| Skills | Skills + Subagents + Hooks (e.g. XO) | |
| Setup effort | Low — Markdown files in .claude/skills/ | Higher, as it requires Node.js, scripts, and agent definitions |
| Task suitability | Best for linear, single-purpose workflows (e.g., standard SQL formatting). | Best for multi-phase workflows with strict phase dependencies. |
| Enforcement | Advisory, the model reads and follows but can deviate | Structural due to tool denial and event-driven scripts |
| Context management | Relies on the model’s context window | Survives compaction via hook-managed state |
| Quality assurance | Same agent authors and reviews | High; separate agents for authoring and review |
| Maintenance | Update Markdown when conventions change | More moving parts to update and maintain. Markdowns for skills, agent definitions, package dependencies. |
| Which to use? | Experimenting, working on isolated tasks, or simply requiring agents to remember a specific set of rules or documentation patterns. | The workflow you are building has dependencies (e.g. Step A must validate before Step B runs), requires integrations with external APIs/MCPs, or needs to be cleanly packaged and distributed across your data engineering team |
For most teams starting out, skills alone are the right choice. They cover the highest-frequency gaps with the least friction. The plugin approach becomes worth the overhead when tasks regularly exceed what a single agent context window can hold, or when the cost of an unreviewed error is high enough to justify separating authoring and review.
How Mantel can help
Our award-winning Snowflake team is constantly evolving how we use CoCo. One of our most popular ways to add value is to conduct a fact-based review of an existing Data Delivery process and then set up a COE to measurably accelerate it.
We’d love to help you unlock what’s possible.
